An Introduction to Reinforcement Learning

Size: px

Start display at page:

Download "An Introduction to Reinforcement Learning"

Maria Flowers
5 years ago
Views:

1 An Introduction to Reinforcement Learning Shivaram Kalyanakrishnan Department of Computer Science and Engineering Indian Institute of Technology Bombay April 2018

2 What is Reinforcement Learning? Shivaram Kalyanakrishnan 1/21

3 What is Reinforcement Learning? [Video 1 of toddler learning to walk] 1. Shivaram Kalyanakrishnan 1/21

4 What is Reinforcement Learning? [Video 1 of toddler learning to walk] 1. Shivaram Kalyanakrishnan 1/21

5 What is Reinforcement Learning? [Video 1 of toddler learning to walk] Learning to Drive a Bicycle using Reinforcement Learning and Shaping Jette Randløv and Preben Alstrøm. ICML Shivaram Kalyanakrishnan 1/21

6 What is Reinforcement Learning? [Video 1 of toddler learning to walk] Learning to Drive a Bicycle using Reinforcement Learning and Shaping Jette Randløv and Preben Alstrøm. ICML Learning by trial and error to perform sequential decision making Shivaram Kalyanakrishnan 1/21

7 Our View of Reinforcement Learning Operations Research (Dynamic Programming) Control Theory Psychology (Animal Behaviour) Reinforcement Learning Neuroscience Artificial Intelligence and Computer Science Shivaram Kalyanakrishnan 2/21

8 Our View of Reinforcement Learning Operations Research (Dynamic Programming) Control Theory Psychology (Animal Behaviour) Reinforcement Learning Neuroscience Artificial Intelligence and Computer Science B. F. Skinner Shivaram Kalyanakrishnan 2/21

9 Our View of Reinforcement Learning R. E. Bellman Operations Research (Dynamic Programming) Control Theory Psychology (Animal Behaviour) Reinforcement Learning Neuroscience Artificial Intelligence and Computer Science B. F. Skinner Shivaram Kalyanakrishnan 2/21

10 Our View of Reinforcement Learning R. E. Bellman Operations Research (Dynamic Programming) Control Theory D. P. Bertsekas Psychology (Animal Behaviour) Reinforcement Learning Neuroscience Artificial Intelligence and Computer Science B. F. Skinner Shivaram Kalyanakrishnan 2/21

11 Our View of Reinforcement Learning R. E. Bellman Operations Research (Dynamic Programming) Control Theory D. P. Bertsekas Psychology (Animal Behaviour) Reinforcement Learning Neuroscience Artificial Intelligence and Computer Science B. F. Skinner W. Schultz Shivaram Kalyanakrishnan 2/21

12 Our View of Reinforcement Learning R. E. Bellman Operations Research (Dynamic Programming) Control Theory D. P. Bertsekas Psychology (Animal Behaviour) Reinforcement Learning Neuroscience Artificial Intelligence and Computer Science B. F. Skinner W. Schultz R. S. Sutton Shivaram Kalyanakrishnan 2/21

13 Resources Reinforcement Learning: A Survey. Leslie Pack Kaelbling, Michael L. Littman, and Andrew W. Moore. JAIR Shivaram Kalyanakrishnan 3/21

Reinforcement Learning: An Introduction Richard S. Sutton and Andrew G. Barto.

14 Resources Reinforcement Learning: A Survey. Leslie Pack Kaelbling, Michael L. Littman, and Andrew W. Moore. JAIR Reinforcement Learning: An Introduction Richard S. Sutton and Andrew G. Barto. MIT Press, ( draft also now on-line). Algorithms for Reinforcement Learning Csaba Szepesvári. Morgan & Claypool, Shivaram Kalyanakrishnan 3/21

15 Today s Class 1. Markov Decision Problems 2. Planning and learning 3. Deep Reinforcement Learning 4. Summary Shivaram Kalyanakrishnan 4/21

16 Today s Class 1. Markov Decision Problems 2. Planning and learning 3. Deep Reinforcement Learning 4. Summary Shivaram Kalyanakrishnan 4/21

17 Markov Decision Problem s t+1 r t+1 ENVIRONMENT T R state s t reward r t LEARNING AGENT action a t π : S A S: set of states. A: set of actions. T : transition function. s S, a A, T (s, a) is a distribution over S. R: reward function. s, s S, a A, R(s, a, s ) is a finite real number. γ: discount factor. 0 γ < 1. Shivaram Kalyanakrishnan 5/21

18 Markov Decision Problem s t+1 r t+1 ENVIRONMENT T R state s t reward r t LEARNING AGENT action a t π : S A S: set of states. A: set of actions. T : transition function. s S, a A, T (s, a) is a distribution over S. R: reward function. s, s S, a A, R(s, a, s ) is a finite real number. γ: discount factor. 0 γ < 1. Trajectory over time: s 0, a 0, r 0, s 1, a 1, r 1,..., s t, a t, r t, s t+1,.... Shivaram Kalyanakrishnan 5/21

19 Markov Decision Problem s t+1 r t+1 ENVIRONMENT T R state s t reward r t LEARNING AGENT action a t π : S A S: set of states. A: set of actions. T : transition function. s S, a A, T (s, a) is a distribution over S. R: reward function. s, s S, a A, R(s, a, s ) is a finite real number. γ: discount factor. 0 γ < 1. Trajectory over time: s 0, a 0, r 0, s 1, a 1, r 1,..., s t, a t, r t, s t+1,.... Value, or expected long-term reward, of state s under policy π: V π (s) = E[r 0 + γr 1 + γ 2 r to s 0 = s, a i = π(s i )]. Shivaram Kalyanakrishnan 5/21

20 Markov Decision Problem s t+1 r t+1 ENVIRONMENT T R state s t reward r t LEARNING AGENT action a t π : S A S: set of states. A: set of actions. T : transition function. s S, a A, T (s, a) is a distribution over S. R: reward function. s, s S, a A, R(s, a, s ) is a finite real number. γ: discount factor. 0 γ < 1. Trajectory over time: s 0, a 0, r 0, s 1, a 1, r 1,..., s t, a t, r t, s t+1,.... Value, or expected long-term reward, of state s under policy π: V π (s) = E[r 0 + γr 1 + γ 2 r to s 0 = s, a i = π(s i )]. Objective: Find π such that V π (s) is maximal s S. Shivaram Kalyanakrishnan 5/21

21 Examples What are the agent and environment? What are S, A, T, and R? Shivaram Kalyanakrishnan 6/21

22 Examples What are the agent and environment? What are S, A, T, and R? 1. Shivaram Kalyanakrishnan 6/21

23 Examples What are the agent and environment? What are S, A, T, and R? An Application of Reinforcement Learning to Aerobatic Helicopter Flight Pieter Abbeel, Adam Coates, Morgan Quigley, and Andrew Y. Ng. NIPS SH-3-Sea-King-helicopter-191.preview.jpg Shivaram Kalyanakrishnan 6/21

Morgan Quigley, and Andrew Y. Ng. NIPS 2006. 1. http://www.chess-game-strategies.

24 Examples What are the agent and environment? What are S, A, T, and R? [Video 3 of Tetris] An Application of Reinforcement Learning to Aerobatic Helicopter Flight Pieter Abbeel, Adam Coates, Morgan Quigley, and Andrew Y. Ng. NIPS SH-3-Sea-King-helicopter-191.preview.jpg 3. Shivaram Kalyanakrishnan 6/21

25 Today s Class 1. Markov decision problems 2. Planning and learning 3. Deep Reinforcement Learning 4. Summary Shivaram Kalyanakrishnan 7/21

26 Bellman s Equations Recall that V π (s) = E[r 0 + γr 1 + γ 2 r s 0 = s, a i = π(s i )]. Bellman s Equations ( s S): V π (s) = s S T (s, π(s), s ) [R(s, π(s), s ) + γv π (s )]. V π is called the value function of π. Shivaram Kalyanakrishnan 8/21

27 Bellman s Equations Recall that V π (s) = E[r 0 + γr 1 + γ 2 r s 0 = s, a i = π(s i )]. Bellman s Equations ( s S): V π (s) = s S T (s, π(s), s ) [R(s, π(s), s ) + γv π (s )]. V π is called the value function of π. Define ( s S, a A): Q π (s, a) = s S T (s, a, s ) [R(s, a, s ) + γv π (s )]. Q π is called the action value function of π. V π (s) = Q π (s, π(s)). Shivaram Kalyanakrishnan 8/21

28 Bellman s Equations Recall that V π (s) = E[r 0 + γr 1 + γ 2 r s 0 = s, a i = π(s i )]. Bellman s Equations ( s S): V π (s) = s S T (s, π(s), s ) [R(s, π(s), s ) + γv π (s )]. V π is called the value function of π. Define ( s S, a A): Q π (s, a) = s S T (s, a, s ) [R(s, a, s ) + γv π (s )]. Q π is called the action value function of π. V π (s) = Q π (s, π(s)). The variables in Bellman s Equations are the V π (s). S linear equations in S unknowns. Shivaram Kalyanakrishnan 8/21

29 Bellman s Equations Recall that V π (s) = E[r 0 + γr 1 + γ 2 r s 0 = s, a i = π(s i )]. Bellman s Equations ( s S): V π (s) = s S T (s, π(s), s ) [R(s, π(s), s ) + γv π (s )]. V π is called the value function of π. Define ( s S, a A): Q π (s, a) = s S T (s, a, s ) [R(s, a, s ) + γv π (s )]. Q π is called the action value function of π. V π (s) = Q π (s, π(s)). The variables in Bellman s Equations are the V π (s). S linear equations in S unknowns. Thus, given S, A, T, R, γ, and a fixed policy π, we can solve Bellman s equations efficiently to obtain, s S, a A, V π (s) and Q π (s, a). Shivaram Kalyanakrishnan 8/21

30 Bellman s Optimality Equations Let Π be the set of all policies. What is its cardinality? Shivaram Kalyanakrishnan 9/21

31 Bellman s Optimality Equations Let Π be the set of all policies. What is its cardinality? It can be shown that there exists a policy π Π such that π Π s S: V π (s) V π (s). V π is denoted V, and Q π is denoted Q. There could be multiple optimal policies π, but V and Q are unique. Shivaram Kalyanakrishnan 9/21

32 Bellman s Optimality Equations Let Π be the set of all policies. What is its cardinality? It can be shown that there exists a policy π Π such that π Π s S: V π (s) V π (s). V π is denoted V, and Q π is denoted Q. There could be multiple optimal policies π, but V and Q are unique. Bellman s Optimality Equations ( s S): V (s) = max a A s S T (s, a, s ) [R(s, a, s ) + γv (s )]. Shivaram Kalyanakrishnan 9/21

33 Bellman s Optimality Equations Let Π be the set of all policies. What is its cardinality? It can be shown that there exists a policy π Π such that π Π s S: V π (s) V π (s). V π is denoted V, and Q π is denoted Q. There could be multiple optimal policies π, but V and Q are unique. Bellman s Optimality Equations ( s S): V (s) = max a A s S T (s, a, s ) [R(s, a, s ) + γv (s )]. V can be computed up to arbitrary precision using Value Iteration. Shivaram Kalyanakrishnan 9/21

34 Planning and Learning Planning problem: Given S, A, T, R, γ, how can we find an optimal policy π? We need to be computationally efficient. Shivaram Kalyanakrishnan 10/21

35 Planning and Learning Planning problem: Given S, A, T, R, γ, how can we find an optimal policy π? We need to be computationally efficient. Learning problem: Given S, A, γ, and the facility to follow a trajectory by sampling from T and R, how can we find an optimal policy π? We need to be sampleefficient. Shivaram Kalyanakrishnan 10/21

36 The Learning Problem Shivaram Kalyanakrishnan 11/21

37 The Learning Problem Shivaram Kalyanakrishnan 11/21

38 The Learning Problem Shivaram Kalyanakrishnan 11/21

39 The Learning Problem Shivaram Kalyanakrishnan 11/21

40 The Learning Problem Shivaram Kalyanakrishnan 11/21

41 The Learning Problem Shivaram Kalyanakrishnan 11/21

42 The Learning Problem r 0 = 1. Shivaram Kalyanakrishnan 11/21

43 The Learning Problem r 0 = 1. r 1 = 2. Shivaram Kalyanakrishnan 11/21

44 The Learning Problem r 0 = 1. r 1 = 2. r 2 = 10. Shivaram Kalyanakrishnan 11/21

45 The Learning Problem r 0 = 1. r 1 = 2. r 2 = 10. How to take actions so as to maximise expected long-term reward E[r 0 + γr 1 + γ 2 r ]? Shivaram Kalyanakrishnan 11/21

46 The Learning Problem r 0 = 1. r 1 = 2. r 2 = 10. How to take actions so as to maximise expected long-term reward E[r 0 + γr 1 + γ 2 r ]? [Note that there exists an (unknown) optimal policy.] Shivaram Kalyanakrishnan 11/21

47 Q-Learning Keep a running estimate of the expected long-term reward obtained by taking each action from each state s, and acting optimally thereafter. Q red green Shivaram Kalyanakrishnan 12/21

48 Q-Learning Keep a running estimate of the expected long-term reward obtained by taking each action from each state s, and acting optimally thereafter. Q red green Update these estimates based on experience (s t, a t, r t, s t+1 ): Q(s t, a t ) Q(s t, a t ) + α t{r t + max a Q(s t+1, a) Q(s t, a t )}. Shivaram Kalyanakrishnan 12/21

49 Q-Learning Keep a running estimate of the expected long-term reward obtained by taking each action from each state s, and acting optimally thereafter. Q red green Update these estimates based on experience (s t, a t, r t, s t+1 ): Q(s t, a t ) Q(s t, a t ) + α t{r t + max a Q(s t+1, a) Q(s t, a t )}. Shivaram Kalyanakrishnan 12/21

50 Q-Learning Keep a running estimate of the expected long-term reward obtained by taking each action from each state s, and acting optimally thereafter. Q red green Update these estimates based on experience (s t, a t, r t, s t+1 ): Q(s t, a t ) Q(s t, a t ) + α t{r t + max a Q(s t+1, a) Q(s t, a t )}. Shivaram Kalyanakrishnan 12/21

51 Q-Learning Keep a running estimate of the expected long-term reward obtained by taking each action from each state s, and acting optimally thereafter. Q red green Update these estimates based on experience (s t, a t, r t, s t+1 ): Q(s t, a t ) Q(s t, a t ) + α t{r t + max a Q(s t+1, a) Q(s t, a t )}. Act greedily based on the estimates (exploit) most of the time, but still Make sure to explore each action enough times. Shivaram Kalyanakrishnan 12/21

52 Q-Learning Keep a running estimate of the expected long-term reward obtained by taking each action from each state s, and acting optimally thereafter. Q red green Update these estimates based on experience (s t, a t, r t, s t+1 ): Q(s t, a t ) Q(s t, a t ) + α t{r t + max a Q(s t+1, a) Q(s t, a t )}. Act greedily based on the estimates (exploit) most of the time, but still Make sure to explore each action enough times. Q-learning will converge and induce an optimal policy! Shivaram Kalyanakrishnan 12/21

53 Q-Learning Algorithm Let Q be our guess of Q : for every state s and action a, initialise Q(s, a) arbitrarily. We will start in some state s 0. For t = 0, 1, 2,... Take an action a t, chosen uniformly at random with probability ɛ, and to be argmax a Q(s t, a) with probability 1 ɛ. The environment will generate next state s t+1 and reward r t+1. Update: Q(s t, a t ) Q(s t, a t ) + α t (r t+1 + γ max a A Q(s t+1, a) Q(s t, a t )). [ɛ: parameter for ɛ-greedy exploration] [α t: learning rate] [r t+1 +γ max a A Q(s t+1, a) Q(s t, a t ): temporal difference prediction error] Shivaram Kalyanakrishnan 13/21

54 Q-Learning Algorithm Let Q be our guess of Q : for every state s and action a, initialise Q(s, a) arbitrarily. We will start in some state s 0. For t = 0, 1, 2,... Take an action a t, chosen uniformly at random with probability ɛ, and to be argmax a Q(s t, a) with probability 1 ɛ. The environment will generate next state s t+1 and reward r t+1. Update: Q(s t, a t ) Q(s t, a t ) + α t (r t+1 + γ max a A Q(s t+1, a) Q(s t, a t )). [ɛ: parameter for ɛ-greedy exploration] [α t: learning rate] [r t+1 +γ max a A Q(s t+1, a) Q(s t, a t ): temporal difference prediction error] For ɛ (0, 1] and α t = 1 t, it can be proven that as t, Q Q. Q-Learning Christopher J. C. H. Watkins and Peter Dayan. Machine Learning, Shivaram Kalyanakrishnan 13/21

55 Practice In Spite of the Theory! Task State State Policy Representation Aliasing Space (Number of features) Backgammon (T1992) Absent Discrete Neural network (198) Job-shop scheduling (ZD1995) Absent Discrete Neural network (20) Tetris (BT1906) Absent Discrete Linear (22) Elevator dispatching (CB1996) Present Continuous Neural network (46) Acrobot control (S1996) Absent Continuous Tile coding (4) Dynamic channel allocation (SB1997) Absent Discrete Linear (100 s) Active guidance of finless rocket (GM2003) Present Continuous Neural network (14) Fast quadrupedal locomotion (KS2004) Present Continuous Parameterized policy (12) Robot sensing strategy (KF2004) Present Continuous Linear (36) Helicopter control (NKJS2004) Present Continuous Neural network (10) Dynamic bipedal locomotion (TZS2004) Present Continuous Feedback control policy (2) Adaptive job routing/scheduling (WS2004) Present Discrete Tabular (4) Robot soccer keepaway (SSK2005) Present Continuous Tile coding (13) Robot obstacle negotiation (LSYSN2006) Present Continuous Linear (10) Optimized trade execution (NFK2007) Present Discrete Tabular (2-5) Blimp control (RPHB2007) Present Continuous Gaussian Process (2) 9 9 Go (SSM2007) Absent Discrete Linear ( 1.5 million) Ms. Pac-Man (SL2007) Absent Discrete Rule list (10) Autonomic resource allocation (TJDB2007) Present Continuous Neural network (2) General game playing (FB2008) Absent Discrete Tabular (part of state space) Soccer opponent hassling (GRT2009) Present Continuous Neural network (9) Adaptive epilepsy treatment (GVAP2008) Present Continuous Extremely rand. trees (114) Computer memory scheduling (IMMC2008) Absent Discrete Tile coding (6) Motor skills (PS2008) Present Continuous Motor primitive coeff. (100 s) Combustion Control (HNGK2009) Present Continuous Parameterized policy (2-3) Shivaram Kalyanakrishnan 14/21

56 Practice In Spite of the Theory! Task State State Policy Representation Aliasing Space (Number of features) Backgammon (T1992) Absent Discrete Neural network (198) Job-shop scheduling (ZD1995) Absent Discrete Neural network (20) Tetris (BT1906) Absent Discrete Linear (22) Elevator dispatching (CB1996) Present Continuous Neural network (46) Acrobot control (S1996) Absent Continuous Tile coding (4) Dynamic channel allocation (SB1997) Absent Discrete Linear (100 s) Active guidance of finless rocket (GM2003) Present Continuous Neural network (14) Fast quadrupedal locomotion (KS2004) Present Continuous Parameterized policy (12) Robot sensing strategy (KF2004) Present Continuous Linear (36) Helicopter control (NKJS2004) Present Continuous Neural network (10) Dynamic bipedal locomotion (TZS2004) Present Continuous Feedback control policy (2) Adaptive job routing/scheduling (WS2004) Present Discrete Tabular (4) Robot soccer keepaway (SSK2005) Present Continuous Tile coding (13) Robot obstacle negotiation (LSYSN2006) Present Continuous Linear (10) Optimized trade execution (NFK2007) Present Discrete Tabular (2-5) Blimp control (RPHB2007) Present Continuous Gaussian Process (2) 9 9 Go (SSM2007) Absent Discrete Linear ( 1.5 million) Ms. Pac-Man (SL2007) Absent Discrete Rule list (10) Autonomic resource allocation (TJDB2007) Present Continuous Neural network (2) General game playing (FB2008) Absent Discrete Tabular (part of state space) Soccer opponent hassling (GRT2009) Present Continuous Neural network (9) Adaptive epilepsy treatment (GVAP2008) Present Continuous Extremely rand. trees (114) Computer memory scheduling (IMMC2008) Absent Discrete Tile coding (6) Motor skills (PS2008) Present Continuous Motor primitive coeff. (100 s) Combustion Control (HNGK2009) Present Continuous Parameterized policy (2-3) Shivaram Kalyanakrishnan 14/21

57 Practice In Spite of the Theory! Task State State Policy Representation Aliasing Space (Number of features) Backgammon (T1992) Absent Discrete Neural network (198) Job-shop scheduling (ZD1995) Absent Discrete Neural network (20) Tetris (BT1906) Absent Discrete Linear (22) Elevator dispatching (CB1996) Present Continuous Neural network (46) Acrobot control (S1996) Absent Continuous Tile coding (4) Dynamic channel allocation (SB1997) Absent Discrete Linear (100 s) Active guidance of finless rocket (GM2003) Present Continuous Neural network (14) Fast quadrupedal locomotion (KS2004) Present Continuous Parameterized policy (12) Robot sensing strategy (KF2004) Present Continuous Linear (36) Helicopter control (NKJS2004) Present Continuous Neural network (10) Dynamic bipedal locomotion (TZS2004) Present Continuous Feedback control policy (2) Adaptive job routing/scheduling (WS2004) Present Discrete Tabular (4) Robot soccer keepaway (SSK2005) Present Continuous Tile coding (13) Robot obstacle negotiation (LSYSN2006) Present Continuous Linear (10) Optimized trade execution (NFK2007) Present Discrete Tabular (2-5) Blimp control (RPHB2007) Present Continuous Gaussian Process (2) 9 9 Go (SSM2007) Absent Discrete Linear ( 1.5 million) Ms. Pac-Man (SL2007) Absent Discrete Rule list (10) Autonomic resource allocation (TJDB2007) Present Continuous Neural network (2) General game playing (FB2008) Absent Discrete Tabular (part of state space) Soccer opponent hassling (GRT2009) Present Continuous Neural network (9) Adaptive epilepsy treatment (GVAP2008) Present Continuous Extremely rand. trees (114) Computer memory scheduling (IMMC2008) Absent Discrete Tile coding (6) Motor skills (PS2008) Present Continuous Motor primitive coeff. (100 s) Combustion Control (HNGK2009) Present Continuous Parameterized policy (2-3) Shivaram Kalyanakrishnan 14/21

58 Practice In Spite of the Theory! Task State State Policy Representation Aliasing Space (Number of features) Backgammon (T1992) Absent Discrete Neural network (198) Job-shop scheduling (ZD1995) Absent Discrete Neural network (20) Tetris (BT1906) Absent Discrete Linear (22) Elevator dispatching (CB1996) Present Continuous Neural network (46) Acrobot control (S1996) Absent Continuous Tile coding (4) Dynamic channel allocation (SB1997) Absent Discrete Linear (100 s) Active guidance of finless rocket (GM2003) Present Continuous Neural network (14) Fast quadrupedal locomotion (KS2004) Present Continuous Parameterized policy (12) Robot sensing strategy (KF2004) Present Continuous Linear (36) Helicopter control (NKJS2004) Present Continuous Neural network (10) Dynamic bipedal locomotion (TZS2004) Present Continuous Feedback control policy (2) Adaptive job routing/scheduling (WS2004) Present Discrete Tabular (4) Robot soccer keepaway (SSK2005) Present Continuous Tile coding (13) Robot obstacle negotiation (LSYSN2006) Present Continuous Linear (10) Optimized trade execution (NFK2007) Present Discrete Tabular (2-5) Blimp control (RPHB2007) Present Continuous Gaussian Process (2) 9 9 Go (SSM2007) Absent Discrete Linear ( 1.5 million) Ms. Pac-Man (SL2007) Absent Discrete Rule list (10) Autonomic resource allocation (TJDB2007) Present Continuous Neural network (2) General game playing (FB2008) Absent Discrete Tabular (part of state space) Soccer opponent hassling (GRT2009) Present Continuous Neural network (9) Adaptive epilepsy treatment (GVAP2008) Present Continuous Extremely rand. trees (114) Computer memory scheduling (IMMC2008) Absent Discrete Tile coding (6) Motor skills (PS2008) Present Continuous Motor primitive coeff. (100 s) Combustion Control (HNGK2009) Present Continuous Parameterized policy (2-3) Perfect representations (fully observable, enumerable states) are impractical. Shivaram Kalyanakrishnan 14/21

59 Today s Class 1. Markov Decision Problems 2. Planning and learning 3. Deep Reinforcement Learning 4. Summary Shivaram Kalyanakrishnan 15/21

60 Typical Neural Network-based Representation of Q 1. Shivaram Kalyanakrishnan 16/21

61 ATARI 2600 Games (MKSRVBGRFOPBSAKKWLH2015) [Breakout video 1 ] 1. Shivaram Kalyanakrishnan 17/21

62 ATARI 2600 Games (MKSRVBGRFOPBSAKKWLH2015) [Breakout video 1 ] 1. Shivaram Kalyanakrishnan 17/21

63 AlphaGo (SHMGSDSAPLDGNKSLLKGH2016) March 2016: DeepMind s program beats Go champion Lee Sedol Shivaram Kalyanakrishnan 18/21

AlphaGo (SHMGSDSAPLDGNKSLLKGH2016) 1. http://static1.uk.businessinsider.

64 AlphaGo (SHMGSDSAPLDGNKSLLKGH2016) 1. screen%20shot% %20at%2014.png Shivaram Kalyanakrishnan 18/21

65 Learning Algorithm: Batch Q-learning 1. Represent action value function Q as a neural network. 2. Gather data (on the simulator) by taking ɛ-greedy actions w.r.t. Q: (s 1, a 1, r 1, s 2, a 2, r 2, s 3, a 3, r 3,... s D, a D, r D, s D+1 ). 3. Train the network such that Q(s t, a t) r t + max a Q(s t+1, a). Go to 2. Shivaram Kalyanakrishnan 19/21

66 Learning Algorithm: Batch Q-learning 1. Represent action value function Q as a neural network. AlphaGo: Use both a policy network and an action value network. 2. Gather data (on the simulator) by taking ɛ-greedy actions w.r.t. Q: (s 1, a 1, r 1, s 2, a 2, r 2, s 3, a 3, r 3,... s D, a D, r D, s D+1 ). AlphaGo: Use Monte Carlo Tree Search for action selection 3. Train the network such that Q(s t, a t) r t + max a Q(s t+1, a). Go to 2. AlphaGo: Trained using self-play. Shivaram Kalyanakrishnan 19/21

67 Today s Class 1. Markov Decision Problems 2. Planning and learning 3. Deep Reinforcement Learning 4. Summary Shivaram Kalyanakrishnan 20/21

68 Summary Learning by trial and error to perform sequential decision making. Do not program behaviour! Rather, specify goals. Rich history, at confluence of several fields of study, firm foundation. Given an MDP (S, A, T, R, γ), we have to find a policy π : S A that yields high expected long-term reward from states. An optimal value function V exists, and it induces an optimal policy π (several optimal policies might exist). Under planning, we are given S, A, T, R, and γ. We may compute V and π using a dynamic programming algorithm such as policy iteration. In the learning context, we are given S, A, and γ: we may sample T and R in a sequential manner. We can still converge to V and π by applying a temporal difference learning method such as Q-learning. Limited in practice by quality of the representation used. Deep neural networks address the representation problem in some domains, and have yielded impressive results. Shivaram Kalyanakrishnan 21/21

An Introduction to Reinforcement Learning

An Introduction to Reinforcement Learning Shivaram Kalyanakrishnan shivaram@csa.iisc.ernet.in Department of Computer Science and Automation Indian Institute of Science August 2014 What is Reinforcement